- Soon Kai Hong
What Makes The AMD Radeon RX 7900 XTX Tick?
We've certainly yet to get our hands on the latest graphics card from AMD which is all so looking so promising. Honestly, we wished we could test it out right now! But perhaps before that, let's actually dive in deeper and take a look at what the new Radeon GPUs really have to offer and what is it that makes them amazing.
Back in Vegas, we had the opportunity to sit down with the professionals in charge of all the work that goes behind the new Radeon 7000 series and we seriously learnt a lot. Too much in fact. To the point that if we were to share every single thing they talked about, you're probably going to be here for at least 3 hours. So we'll keep it simple.
There are a few things to talk about here and we'll start with RDNA 3 in itself.
Chiplet Design
The key focus for RDNA 3 and arguably the most iconic by far has to be the use of the chiplet design. AMD mentioned that they've actually been working on the idea for quite a few years now with the idea spawning from a simple notepad in a hotel room! But anyways, what is a chiplet design?
The chiplet design isn't anything new per se and has already been seen in the majority of server-grade CPUs such as the EPYC series from AMD themselves. With a standard monolithic design, chipmakers such as TSMC have to actually shrink the entire chip based on the selected nanometer process and that has to be so even if certain blocks within the chip itself aren't updated and or won't work differently from the shrinking process. In addition, if one chip out of the entire silicon wafer does not make the minimum threshold for the intended SKU or performance target, that entire chip has to be binned into a lower-tier segment as a product. Now there are benefits for going with a monolithic design as opposed to a chiplet design but we won't delve in too much here.
Let's focus on the chiplet design and what it brings to the table.
The first is efficiency in yields which directly translate to cost savings. Unlike a monolithic design, even if there's a defective or not up to standard chiplet among the bunch, all that needs to be done is to simply swap it out for another chiplet, thereby saving the need to discard the entire chip in itself. Additionally, due to the inherent nature of the chiplet design, production can also be optimised by using different process nodes for different parts of the chiplet such as the main core, the memory controller, the I/O, so on and so forth.
Taking the Radeon 7000 series as an example, AMD went with TSMC's latest 5nm process node for the Graphics Compute Die (GCD) itself but went with 6nm for the Memory Controller Die (MCD). Why is that so? Well, MCDs don't benefit too much from a shrinking in die and the cost to performance ratio doesn't make sense. But perhaps more importantly, a chiplet design allows AMD to mix and match to achieve the best of both worlds and that is exactly the path that AMD has taken.
According to AMD, density gains are diminishing yet costs are increasing exponentially. Thus the chiplet design was the way to go, providing pretty much a linear cost according to core count.
It Isn't Magical
But doesn't mean that going with a chiplet design is a magical transition at the snap of the finger. There are many things to address when implementing said design for GPUs.
Unlike in an EPYC chip, the interface between the different chiplets simply require 100's of signals. On a GPU however, that number increases to a staggering 10's or 1000's of signals. In order to combat this, AMD has to design a whole new interface to be able to handle that amount of bandwidth. This is where the High Performance Fanout and Infinity Links come into play.
Infinity Links operate at 9.2Gb/s with High Performance Fanout providing almost 10x the bandwidth density as compared to the now standard Infinity Fabric On-Package links used in Ryzen and EPYC. This is what enables AMD to claim an industry-leading peak bandwidth of 5.3TB/s. But for those of you who know your stuff, latency is still an issue to tackle. Compared to a standard monolithic design, even with the Infinity Link, a chiplet design still costs a moderate amount of latency. Thus, AMD eliminates this by going with +43% on the base Infinity Fabric Clock and +18% on the Graphics Game Clock.
The result of all these hard work is up to a 54% increase in performance per watt against RDNA 2. At a glance, this is what you can expect from the new Radeon 7000 series such as the Radeon RX 7900 XTX.
Now we can go on and on and talk way more technical things but that's not our expertise as of yet. Instead, let's bring it back down to the basics. With such advancements in technology and the use of the chiplet design, what can we expect in terms of actual real world performance?
Performance
Here are the specs at a glance.
The Radeon RX 7900 XTX is the flagship, featuring 96 RDNA 3 CUs for a total of 6,144 Stream Processors. It runs at a Game Clock of 2.3GHz with a Boost Clock of up to 2.5GHz and has 96MB of Infinity Cache alongside 24GB of GDDR6 in a 384-bit bus running at 20Gbps.
The Radeon RX 7900 XT is no slouch either, featuring 84 RDNA 3 CUs for a total of 5,376 Stream Processors. It runs at a Game Clock of 2GHz with a Boost Clock of up to 2.4GHz and has 80MB of Infinity Cache alongside 20GB of GDDR6 in a 320-bit bus running at 20Gbps.
At Native 4K settings, the RX 7900 XTX was able to achieve almost 67% more performance in comparison to the previous flagship RX 6950 XT in popular titles such as Cyberpunk 2077 and the latest Call of Duty: Modern Warfare 2. We're looking at 72 fps vs 43 fps in Cyberpunk 2077 and 139 fps and 92 fps in Call of Duty: Modern Warfare 2 respectively.
If ray-tracing is to be in the picture at 4K, performance is now even better especially with FSR 2.0 or above enabled.
But perhaps the most interesting capabilities of the new card is the fact that it can do 8K gaming surprisingly well. Going just by AMD's internal testing, you'll be able to game at 8K and achieve extremely playable framerates with Uncharted: Legacy of Thieves Collection and Assassin's Creed Valhalla. The catch however? FSR must be enabled for sure and the Ultrawide setting is basically half that of 8K. So the 8K Ultrawide benchmark isn't truly 8K per se. BGut of course, take everything with a grain of salt for we need to grab our hands on the new GPU to really come to any conclusive findings.
Again for those interested, the AMD Radeon RX 7900 XTX will retail for US$999 while the Radeon RX 7900 XT will retail just a little less at US$899. Both of which will be available come 13 December, 2022.
Here's hoping we get our hands on the cards really soon!
The new Radeon 7000 series GPUs from AMD utilises the chiplet design for added efficiency and performance while reducing manufacturing costs, allowing AMD to target the magical US$1,000 mark.
Infinity Links and High Performance Fanout are two new key technologies to address the issue of latency by going with a chiplet design and AMD has managed to instead achieve up to 54% increase in performance with RDNA 3 in comparison to RDNA 2.
At Native 4K settings, the Radeon RX 7900 XTX was able too achieve almost 67% more performance as compared to the Radeon RX 6950 XT.